Someone on HardOCP forums
pointed out that the real-world bandwidth of Vega FE is probably limited by the Synopsys memory controller they are using for HBM. Testing from PCGH.de provided only a bit over 300 GB/sec of bandwidth in testing, despite the fact that the stacks should be able to put through 480 GB/sec at the clock rate they are run at. The Synopsys HBM controller supports only up to 307 GB/sec bandwidth.
This may be part of the explanation for why Vega underperforms so badly in gaming. They were originally going for 512 GB/sec memory bandwidth, but ended up with less than 60% of that. Whatever simulations and other stuff they ran assumed nearly twice as much memory bandwidth as they actually got. Hynix/Micron/Samsung/whoever they're getting HBM2 from underperformed, and apparently Synopsys, from whom they're licensing their memory controller, underperformed as well (or they didn't read the specs carefully enough?) This is one of the drawbacks of running an operation on a small R&D budget: you have to rely on third-party vendors and can't always control the quality of their work.